神经网络对新增样本的学习算法

THE LEARNING ALGORITHM FOR THE STUDY OF NEW PATTERN IN NEURAL NETWORK

对新增样本的快速学习而又不损失原有样本的记忆，是自适应在线系统的要求．本文提出了一种基于对节点激励函数线性化的逐层优化学习算法，为防止由于线性化而造成较大的误差，在损失函数中加入了惩罚项．该算法在每次迭代中，权值矩阵可以显式表达出来．

Learning new sample quickly without degrading the recall of old samples is the requirement of adaptive on line system. In this paper, a new learning procedure is presented which is based on the linearization of neuron activation function. To avoid the big error resulted from the linearization, the penalty terms are added to the cost function. In the course of learning by the algorithm, the optimal solution per interation can be clearly expressed. Computer simulation results indicate the proposed algorithm is feasible and effective.